Online Signature Verification Based on Recurrent Attentional Time-Delay Neural Networks

Xirali Ablat, Qixiang Li, Nurbiya Yadikar, Kurban Ubul

Abstract: Online signature verification, an important biometric authentication technique, faces significant challenges in dealing with the inherent variability of individual handwritten signatures. To solve this problem, we propose an online signature verification method that combines Soft Dynamic Time Warping (Soft-DTW) and deep learning, aiming to improve the accuracy of the system in signature verification tasks. Specifically, we design the attentive time-delay neural network (TGRUN), which is a combination of time-delay neural network (TDNN) and gated recurrent unit (GRU) as well as the attention mechanism. By combining the time-delay property of TDNN and the recurrent capability of GRU, and using the attention mechanism to weight the key time points, the TGRUN model, when processing online signature data can provide more accurate and robust performance in complex environments. In addition, we use Bayesian parameterization combined with gradient optimization to adaptively tune the smoothing parameter \(\gamma \) of Soft-DTW to achieve more accurate sequence matching. The average equal error rates of this research method on MCYT-100, SVC2004 Task2 and self-constructed multilingual datasets are 1.76%, 3.29% and 1.52%. The experimental results show that our method is more advantageous than existing techniques that use manual extraction or learned feature characterization.